HCSE-Net: A Hybrid BiLSTM Framework with Structural Feature Fusion for Tweet Sentiment Classification

The rapid evolution of social media platforms has transformed how individuals communicate and express opinions online. Twitter has gained significant popularity as a source of real-time popular mood because of the nature of tweets, where most of the content is brief, and the user provides a lot of participation. Consequently, the Natural Language Processing (NLP) task aimed at automatically identifying the emotional polarity of tweets. Sentiment analysis makes it possible to reveal significant knowledge about a huge amount of textual data that otherwise would be hard to assess manually [1]. With the increased use of social media in communication and information exchange, automated sentiment analysis has become a valuable component of the current data-driven decision-making system in businesses and research in the past few years, as people have learned how effective and valuable it is to analyse the sentiments on Twitter in order to learn more about the opinion of the audience and maintain their reputation and predict what new social trends will occur. Most tweets are brief, informal, and full of expressive features such as emojis, hashtags, abbreviations, and slang. This is unlike news reports and official paperwork. These attributes give both possibilities and difficulties to the sentiment classification processes. Tweets offer instant and real-time feedback by the users; however, since they are not structured, it is difficult to comprehend the meaning of their tweets [2]. As a result, it has been proposed that the development of useful computer algorithms to detect the sentiment of a tweet has become a major research agenda. Accurate sentiment detection provides valuable insights for business analytics, political forecasting, healthcare monitoring, and crisis management. Firms tend to use customer tweets to check the performance of their products and how to market it. Policymakers analyze sentiment trends to determine how the citizens of the country are feeling about the government projects or social activities. Moreover, in case of crisis or other disasters, the mood analysis of the people can assist the government to locate false data, emotional distress, or genuine concerns of people [3]. These uses indicate that emotion recognition is not just a technique issue, but a useful tool to make informed decisions in a variety of real-life situations.

The reason is that Tweet sentiment analysis is significant; however, it contains numerous issues that distinguish it from other text categorisation tasks. The tweets are usually full of loud language, bad syntax and statements that are dependent on the situation, which may deceive the average algorithms. Implicit attitude, sarcasm, and irony make it more difficult to interpret what an individual means and therefore models need to locate unobtrusive links between words in the context. Additionally, on the internet, language is a constantly evolving phenomenon, so new terms, memes, and acronyms are generated continuously. This renders fixed lexicons or rule systems useless as time goes by Rosenthal et al. [4]. Overcoming these issues requires sophisticated modelling techniques that are able to comprehend local word-level patterns as well as global context linkages.

The vast majority of initial sentiment analysis systems were constructed based on lexicons and using conventional machine learning (ML) techniques. Lexicon-based methods give ratings of the word or sentence with respect to its polarity based on existing emotion dictionaries, giving an interpretable result, but with low flexibility to contextual variations. Some examples of machine learning models include Naive Bayes, Support Vector Machine (SVM) and Decision Trees, which enhance the accuracy of classification by identifying statistical patterns within the data that has been labelled. However, the latter techniques heavily rely on the structural aspects and do not consider sequential relations inherent in language [5]. The need to generate representations of features from raw text using deep learning (DL) algorithms emerged when researchers started to study how to generate representations of features and enhance understanding of the resulting representations as the amount and complexity of social media data grew. Development of deep neural networks, including Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), has contributed greatly to the research in sentiment analysis. CNNs are effective at discovering local semantics, such as phrases or n-grams, but Long Short-Term Memory (LSTM) networks are more effective at modelling the time-connectivity of word sequences. Bidirectional Long Short-Term Memory (BiLSTM) architectures also perform sentiment prediction with more accuracy because they use both forward and backward analysis of text to build a stronger contextual understanding. Recent literature has shown that hybrid architectures with CNN and BiLSTM layers outperform more traditional machine learning algorithms particularly when dealing with short and noisy social media input [6, 7].

Nevertheless, when tweets contain structural information such as hashtags or emoticons which can convey some critical emotional signals, deep learning methods fail. The other emerging trend of sentiment study is the inclusion of additional features and structural data into the neural networks. Basic word embeddings may offer a more detailed view of the underlying emotion, with models typically incorporating meta-linguistic information like punctuations, number of mentions, and usage of emojis. The goal of these hybrid frameworks is to bridge language comprehension and situational awareness, and therefore they are more suitable to real-world Twitter data. With the development of the field, there is the growing interest to produce the systems that are very balanced in terms of performance, interpretability, and computational efficiency [8, 9].

Overall, the tweet sentiment recognition is becoming a subject of study involving machine learning, deep learning, and language analysis. Previous approaches provided the structure upon which automated opinion mining can be established, but the latest deep architectures are reassuring in their ability to deal with the complexity of the language employed on social media. The enduring nature of noise, ambiguity and the constantly changing conditions of the online communication styles highlight the fact that new hybrid models are required that can be effective in obtaining both semantic and structural information. Coupled with these issues, the present-day research work has focused on the creation of innovative architectures that are able to create more contextual knowledge whilst being robust in the context of real-world sentiment analysis in Twitter. Despite significant advancements in deep learning-based sentiment analysis, existing hybrid models primarily focus on semantic embeddings while often ignoring tweet-specific structural sentiment indicators such as emojis, hashtags, punctuation intensity, and informal writing patterns. Although transformer-based architectures provide strong contextual modelling, they introduce substantial computational complexity and deployment overhead for real-time sentiment analytics systems. Therefore, there remains a need for a computationally efficient framework capable of jointly learning contextual and structural sentiment information from noisy Twitter data.

In addition, modern information systems increasingly rely on automated sentiment-aware analytics for decision-support applications, customer feedback mining, crisis management, and public opinion monitoring. The proposed HCSE-Net framework is designed to support such intelligent social media analytics systems through efficient and context-aware sentiment prediction. In this work, the performance and generalizability of a proposed Twitter sentiment analysis system were tested using three publicly available benchmark datasets. The initial one is the Twitter Sentiments Dataset of Hussein [10]. It contains tweets belonging to a number of categories, which have been tagged and their sentiment polarity. The dataset of CS98X Twitter Sentiment Classification was uploaded on Kaggle by Azzopardi [11]. It contains a good combination of tweets that can be employed in supervised sentiment learning. The remote supervision dataset that is famous and was created by Go et al. [12] was also used to offer more diversity and size to the training samples.

To ensure consistency across sources, all datasets underwent preprocessing before testing, including normalisation, tokenisation, and noise reduction techniques. This research categorised sentiment labels into three categories: positive, negative, and neutral. This made it capable of applying the same assessment paradigm to all the datasets. Adding different datasets strengthens the model by subjecting it to differences in style of language, topic distribution and annotation methods. This is used to remove bias in the datasets.

The rest of this paper is intended to outline the research that clearly and logically explains the research that was recommended. Section 1 discusses the key principles of tweet sentiment recognition, including but not restricted to why it matters, what the pitfalls of it are, and what should be done with the work. Section 2 is an intensive scrutiny of the literature that discusses the constraints of the available machine learning methods and deep learning methods, as far as handling informal writing regarding social media is concerned. Section 3 discusses the planned approach, such as architectural design preprocessing techniques, embedding, and the hybrid architecture of deep learning used in the categorisation of sentiment. In Section 4, the experimental setup, the dataset, the evaluation metrics, and an intensive examination of the findings to determine the level at which the HCSE-Net has been achieved are presented. Lastly, Section 5 concludes with the most appropriate findings, identifies the scientific contributions, and suggests potential paths of further study of Twitter sentiment analysis.

The proposed tweet sentiment classification framework is a hybrid deep learning architecture that employs embedding representation, bidirectional sequence modelling, and extra structural characteristics to improve sentiment classification performance, as shown in Figure 1.

The proposed HCSE-Net framework incorporates several auxiliary structural features to enhance sentiment understanding in noisy Twitter text. These features include emoji count, hashtag frequency, punctuation patterns (such as exclamation and question marks), uppercase word ratio, tweet length, repeated character occurrences, and the number of user mentions and URLs. Each structural feature was numerically encoded and normalized using min-max scaling before integration into the network. The extracted auxiliary feature vector was then concatenated with the contextual feature representation generated by the BiLSTM layer through a feature fusion mechanism. This hybrid representation enabled the model to capture both semantic dependencies and tweet-specific structural cues, thereby improving sentiment classification accuracy and robustness on informal social media data, as shown in Figure 2.

Figure 1. Proposed architecture

Figure 2. Sequence wise illustration of proposed methodology

Unlike conventional CNN-BiLSTM and attention-BiLSTM architectures that primarily focus on semantic sequence learning, the proposed HCSE-Net framework jointly models contextual semantic dependencies and tweet-specific structural sentiment indicators within a unified feature fusion architecture. This enables improved handling of noisy and highly informal Twitter content while maintaining lower computational complexity than transformer-based architectures.

Let X denote the input tweet sequence and Wₑ represent the embedding weight matrix. The embedded feature representation E is obtained in Eq. (1):

$E=X \cdot W_{\mathrm{e}}$      (1)

where, E is the embedding output used as input to the BiLSTM network.

The BiLSTM processes the sequence in both forward and backward directions to find dependencies in the context. The LSTM unit's internal gating operations control how information flows through the network.

Forget gate as shown in Eq. (2):

$f_{\mathrm{t}}=\sigma\left(W f \cdot\left\lceil h_{\mathrm{t}-1}, E_{\mathrm{t}}\right\rceil+b f\right)$        (2)

Input gate as shown in Eq. (3):

$i_{\mathrm{t}}=\sigma\left(W i \cdot\left[h_{\mathrm{t}-1}, E_{\mathrm{t}}\right]+b i\right)$        (3) 

Candidate memory state as shown in Eq. (4):

$\hat{\mathrm{C}}_{\mathrm{t}}=\tanh \left(W c \cdot\left\lceil h_{\mathrm{t}-1}, E_{\mathrm{t}}\right\rceil+b c\right)$       (4)

Updated cell state as shown in Eq. (5):

$C_t=f_t \odot C_{t-1}+t \odot \hat{C}_t$       (5)

Hidden output state as shown in Eq. (6):

$h_{\mathrm{t}}=o_{\mathrm{t}} \odot \tanh \left(C_{\mathrm{t}}\right)$       (6)

where, σ represents the sigmoid activation function, ⊙ denotes element-wise multiplication, and hₜ corresponds to the hidden representation at time step t. A feature vector A includes not only sequential features but also extra structural attributes taken from tweets, like the number of emojis, the presence of hashtags, and the patterns of punctuation. We combine these extra features with the BiLSTM output H using a fusion operation as shown in Eq. (7):

$F=[H; A]$      (7)

where, F is the fused feature representation that comes from putting together sequential and structural information.

Then, the fused features go through dense layers to learn more abstract ideas. A softmax classifier is used to figure out the final sentiment prediction as shown in Eq. (8).

$y={softmax}(W d \cdot F+b d)$      (8)

where, y is the probability distribution of the sentiment classes (positive, negative and neutral). They are done using categorical cross-entropy loss and gradient-based optimization techniques to check the optimum model parameter values. The proposed structure will improve its ability to handle noisy and informal language that characterizes Twitter data through the incorporation of a learning sequence in a contextual manner with structural auxiliary encoding. The probability distribution between classes of sentiments (positive, negative and neutral). The proposed outline strengthens the tolerance of the informal and unintelligent language that is characteristic of Twitter data by combining contextual sequence learning with structural auxiliary encoding.

Let $S=\left[s_1, s_2, s_3, \ldots, s_n\right]$ represent the auxiliary structural feature vector extracted from each tweet, where $s_1$ denotes emoji count, $s_2$ hashtag frequency, $s_3$ punctuation intensity, $s_4$ uppercase ratio, $s_5$ tweet length, and other tweet-specific attributes. The extracted features were normalized using min-max normalization as follows in Eq. (9).

$s_i^{\prime}=\frac{s_i-s_{\min}}{s_{\max}-s_{\min}}$       (9)

The normalized auxiliary feature vector was then fused with the contextual representation $h_t$ generated by the BiLSTM network using feature concatenation in Eq. (10).

$F=\left\lceil h_t \oplus S^{\prime}\right\rceil$       (10)

where, F represents the final hybrid feature vector and ⊕ denotes concatenation. This combined representation enabled the proposed HCSE-Net model to simultaneously capture semantic dependencies and structural tweet patterns, improving robustness and sentiment classification accuracy on noisy Twitter data.

The primary novelty of HCSE-Net lies in its lightweight hybrid fusion strategy that combines contextual sequence learning with explicitly engineered tweet-structure sentiment indicators under a unified architecture optimized for noisy social media environments. Unlike transformer-based systems requiring large-scale fine-tuning, HCSE-Net achieves competitive performance with significantly lower computational overhead.

The enhanced effectiveness of the suggested HCSE-Net model compared to the current state-of-the-art methods could be explained by a number of important architectural and design considerations. First, a BiLSTM network was utilised so that the model can consider both forward and backward contextual dependencies in order to obtain a more holistic view of the semantics of tweets as compared to the unidirectional LSTM and CNN-based networks.

Second, contrary to traditional deep learning and hybrid models, such as CNN-LSTM and CNN-BiLSTM, the suggested framework directly includes auxiliary structural characteristics, such as emoji frequency, hashtags, and punctuations. These characteristics are powerful sentiment predictors of social media text and are useful in increasing the capacity of the model to make sense of informal and noisy language.

Thirdly, the fusion mechanism of features successfully merges the semantic representations acquired in the BiLSTM with structural features, which allows the model to learn more discriminatory and richer feature representations. This combination method minimizes loss of information and enhances classification accuracy on all sentiment classes.

Compared to transformer-based models like BERT, Transformer-based models achieve good contextual embeddings, but they must use a high level of computational resources and fine-tuning. The proposed HCSE-Net uses lower computational complexity with competitive and enhanced performance and is thus more efficient and suitable to real-time sentiment analysis applications.

Moreover, the experiment results show that HCSE-Net is more accurate, precise, recalls, and F1-score than MF-CNN-BiLSTM, CNN-LSTM, CNN-BiLSTM, and tuned BERT models. The accelerated convergence in the training loss also suggests that the suggested architecture is able to learn meaningful feature representations.

In general, the bidirectional contextual learning, auxiliary structural feature encoding, and effective feature fusion strategy result in the proposed model HCSE-Net overcoming the main drawbacks of the current approaches, which results in the enhanced robustness, generalization, and classification of noisy Twitter data.

5.1 Error analysis

Although the proposed HCSE-Net framework achieved strong classification performance, several challenging tweet categories still resulted in occasional misclassification. In particular, sarcastic tweets, implicit emotional expressions, abbreviated social media language, and mixed sentiment polarity tweets remained difficult to classify accurately.

For example, tweets such as:

“I absolutely love waiting three hours for customer support.” contain positive lexical indicators but express negative contextual sentiment. Such sarcasm-oriented expressions remain challenging even for advanced hybrid deep learning architectures.

Additionally, tweets with incomplete context, evolving slang, and implicit sentiment occasionally reduced prediction reliability. Future research may integrate transformer-attention mechanisms and sarcasm-aware contextual modelling to further improve classification robustness.

Table 5. Comparison of training time

Table 5 shows that, compared to transformer-based architectures such as BERT, the proposed HCSE-Net framework achieves competitive sentiment classification performance with lower computational complexity and reduced training overhead, making it more suitable for real-time social media analytics systems.